Background-Foreground Segmentation Using Multi-Scale Attention Net (MA-Net): A Deep Learning Approach
Vishruth Boraiah Gowda, Gopalakrishna Madigondanahalli Thimmaiah, Megha Jaishankar, Y. L. Chaitra · Revue d intelligence artificielle · 2023
Background subtraction serves as a critical foundation for numerous computer vision tasks, and a variety of traditional techniques have been proposed.In recent years, deep neural network architectures have emerged as a promising approach, with the UNET architecture being a notable example.However, UNET is considered an older architecture with limitations.To address these limitations, a novel neural network technique called Multi-scale Attention Net (MA-Net) is proposed, which incorporates a self-attention mechanism for adaptively integrating local features with their global dependencies.The attention mechanism within MA-Net enables the capture of complex contextual dependencies.Two distinct blocks are developed for the MA-Net: The Position-wise Attention Block (PAB) and the Multi-scale Fusion Attention Block (MFAB).While PAB models the interdependencies between features from spatial dimensions, representing pixel dependencies in a global view, MFAB capitalizes on fused multi-scale semantic feature fusion to capture channel dependencies between feature maps, effectively segmenting the foreground from the background.The proposed method is evaluated using the CDNET 2014 dataset and demonstrates improved performance under Shadow, dynamic background, and illumination challenges.This study highlights the potential of the MA-Net for advancing the field of background-foreground segmentation in computer vision tasks.